[2/2] Gemini 3: Solving Life’s Messes—But Could It Replace Our Jobs?

So I have some day-to-day uses, like use cases I want to share with you and then see if Gemini 3 can actually do it and how exactly it does it more importantly. So let's say if I want to build a garden shed. I apparently don't know how to do it but I do have a very rough sketch and I also searched a YouTube video 10 minutes long say about how to build the foundation. So can I submit this sketch and the YouTube video link to Gemini 3 and ask it to give me multi-step instructions and probably even suggest the cost budget. Yes, with some caveats. Okay, let's see. So you should make sure that the model has internet access. So generally the models have internet access but sometimes they don't. If they don't have internet access what happens is they'll hallucinate, which is you know the makeup stuff assuming this is what is there in the video. So assuming that they have internet access, in this case most likely it's going to get the transcript of that particular video and because it's YouTube. Yeah, because it's YouTube and you know they own YouTube and so it's pretty easy for them to get that information. And then once they get the transcript then the model will be along with the instructions you passed in. Since it's pretty good at understanding the sketch you should be able to give a pretty good estimate of what is the stuff you need and regarding estimates on how long it'll take you to build and stuff like that. They'll be pretty generic in this case because it's not very clear what your skill level is. So you might probably have to tell how good are you with handling different kinds of tools, right? So if you give it more details you may be able to give a more realistic estimate on those kinds of things. So that for that video piece because they're the same company YouTube and Google and that's why it's a little bit easier for them to access the video. But what about for other models? Let's say does ChatGPT has the same level of access when it comes to YouTube videos? That's a good question. I think in terms for ChatGPT if you give it a video sometimes it says that in its thinking steps it says that it's trying to understand the video. Or think harder. Yes, I think what happens is with some popular videos probably it was part of their training data because they don't tell it for sure. In case of YouTube I'm pretty sure they use it for their data for training their models. But so in my case I had to give it a video and I asked it to summarize the main points in that video and this wasn't a very popular video. So what happened was, this was ChatGPT by the way, so ChatGPT said it was thinking and analyzing watching the video. I was like okay let me see what it's going to do. And then it gave a I think it gave a reasonable explanation of what was happening in the video. And some of the details were wrong. And I found out that it was probably using some details from the transcript or something. But then it wasn't really watching the video or like listening to there's no it wasn't processing the audio. How does it watch the video? Like how does how does the AI model watch something? Oh yeah when you say watch the video it's mostly that they're going to take in a frame of the like multiple frames of the video and then pass that frames as just like we have a they have models for processing an image right. So they have models the models can also process videos as a bunch of frames and you're going to pass all of those frames and each frame is now again you know broken down or a collection of frames are broken down into different kinds of smaller patches and then again they are converted into embeddings which captures in a sense a meaning of what that patch is and you know from that they're going the model is going to process. Yeah so that kind of I was going to ask a tricky question you were talking about to analyze the video sometimes it's based on the transcript that the AI model grabbed. What if it's a silent video it's just a video without any anyone talking or any background information and there's no transcript. So I guess that's exactly how the AI model analyze the video by watching it. Yeah it'll be able to do that and depends on what which company is you know depends on how it has access to these things. You if it is Google's model site they have access to the audio video transcript pretty much everything. So which means they should be able to pretty much process the audio of that video too if necessary. But when it comes to other models let's say OpenAI or Claude or these kinds of models which really don't have access to well those are public videos for sure but I don't know if Google allows them to you know download and use them and things like that. So the yeah the quality of the results depends on what access they have. For these kinds of things especially dealing with the YouTube video and things like that my best bet is always Google's products because they have direct access and the quality of information they have is the best. That makes sense. So on the topic of video or image I have another use case I want to bring up. Let's say I often and I'm sure a lot of people do too have the problem to actually explain my problem. I need help let's say my washing machine there's a part just broke and I don't know the name of the part and I can't describe it. So can I just throw the picture of that broken part to Gemini 3 and just say hey my washing machine doesn't work and seems like this part is where it broke can you help me to fix it? How well does it respond to this type of I can't explain the what the problem was kind of problem? Yeah those kinds of problems you can still give it useful information like the more the more help you give it the more help it can give it back in a sense. So in this case let's say your washing machine is also damaged it has a dent or something like that so it wouldn't be if you just tell upload a picture and tell like hey fix it I could there could be a dent and there could be a problem that it's not actually you know washing clothes for you it'll probably fix the dent for you you'll be like okay there's I see a dent here and you know these are all the things you can you can help with in your case a little bit clarity of yeah some clarity you could be like oh explain the problem like what is more like like if you go to a doctor you just can't tell that I'm not feeling well you know it's like sure then what exactly is happening right it needs a little bit more context he needs some symptoms and you know then he'll be able to figure out what's the most likely illness and things like that so even here you probably have to do the same thing right well not probably you have to do the same thing to get a good answer like tell what exactly is happening like what kind of problem is it able to wash clearly or is it this is a problem like oh the lights are not showing up it's giving weird sounds or if you find that it's the clothes are coming out soapy or like you know it's not really clean and whatnot so once you give this information then you'll be able to help you in those contexts especially um and the more information what I mean by more information is um let's say it's samsung brand right and you also specify the model number in that case or if the model number is available in the picture uh it will probably be able to help and if it has access to internet most likely it will be it might try to look up if there's any known issues for this particular model of of washing machine and maybe they this particular model has an issue and since it uses internet it'll be able to search and then find out what the data is and then it'll tell that oh based on my lookup using search I found that this particular model of samsung has a particular known issue maybe this is related to yours yeah that way to be able to be more that's not exactly like a doctor when trying to like diagnose what your problem is and you have to tell him oh I have a fever I don't feel well on my throat or ear or where wherever that is and um but just compared to the previous model uh gemini 2 or even the older ones um do you think it has a little bit better understanding of the problem or I guess what is the difference and why it's does it live up to the hype um it does but uh not in all applications um like like it is it is a good overall model throughout and it's it's as good as pretty much any other model for all for most of the use cases but people's use cases kind of vary and uh for any generic task uh any simple task you really won't find any difference between this model versus the previous version of the model or any of the models by other providers but as it goes as the task becomes more specific and then more complex and deeper you start to see uh oh how exactly this model is better than the other models I see um curious what is your biggest real life problem you've solved with or you are trying to solve with the current um ai um I well I as part of my work I have to code a lot um so I use uh models for you know writing a lot of code that's my work related uh that's the major application related to work um then I also use these models to uh help me uh give opinions on some things and especially in some cases uh I'd like to have an opinion of multiple models so in in a way it's more like uh uh I give the same question to multiple models uh and then I get all their answers and then I uh I get another model to summarize the results from each of these uh models and then from there I decide like you know which is the one I should spend time on you know looking or reading into so like trying to prioritize things uh yes and also I want to see that because sometimes these models uh one thing is I want to know which model is good at what task uh like not all of them are good at everything so for some specific use cases I want to know which is the better model and other thing is uh sometimes these models kind of hallucinate they make up stuff uh so which means that there is this very low chance that all of them are going to hallucinate on the same problem yeah in some cases yes but not always uh so I want to talk a little bit more about the hallucination um can you explain a little bit more about what exactly that is and how does it affect people's usage with this models yeah hallucination is uh when a model says something that is plausible which could be true but it's not really true can you give us an example um so you uh let's say the model has no access to internet okay and you give it a url and uh and normally the url has some part of the title in there as part of the url right not youtube videos but then if it's like uh url of a blog or something like that generally they have some part of that title if you ask it get me the content of this particular article and then you give the link to that and assuming the model has no internet access this model maybe some models now they tell that hey I don't have internet access or things like that but the older versions of the model they just used to go ahead and create make up stuff and tell like this is what the content is since they know what is that though um that is was it a bug I guess was it a bug or was it kind of intentional because the way the models are trained uh the models are trained uh like these models are auto-regressive uh what I mean by that is they the they decide what what is the next word to tell or next token based on the previous all the previous things which they had said so far um which means that assuming all the things I said so far is true or all the things which the model said so far is true then the next thing is most likely going to be something else but then they may not be true in the real world so it's on the assumption that what so far was said was also true right what so far said might be wrong the model can make mistakes so if it makes a mistake there's a chance that since it's going to use that mistake to tell to for the next words it's going to generate or the next tokens it's more likely that it will make some other mistake and now they now it has two mistakes instead of one and then the third one is probably going to be a mistake because it has enough mistakes there as it keeps on generating more and more text these mistakes kind of accumulate and then it becomes kind of a mess right in a really long context so you see that as the these models are pretty smart on short conversations but if the conversation goes very large very long they kind of you'll notice that oh they're doing some dumb things you know they're not really as smart as they were when the conversation is short yeah and this is what happens in those cases I want to tell you a meme I saw online about this hallucination problem so um there's a there's someone who who are camping or doing whatever outside and then found a mushroom and then he put the mushroom took a picture sent it to whatever model AI model and said hey is it poison can I eat it and the model is like I think it's fine you can eat it and he ate it and what happened is there's a rip right so then somebody put that rip as a photo to the AI has said hey this person this is what he is right now and then the model is like oops I'm sorry so next time I learned that this type of mushroom is poisoning like this this is a type of high hallucination I found it of course it's a meme but I think it shows a lot because it's very important to remind everyone who uses these models on a daily basis especially nowadays I think everyone is using them more or on a daily or weekly basis to always double check because we never really know if it's hallucinating or if it's telling the truth yeah I think in terms of using these models the principle many people won't follow is trust but verify so it's like you can trust the model on a lot of things because there are there are only some specific things which hallucinates on for common knowledge and things like that the models really don't hallucinate these days but people use they just don't use it for any it depends on the amount of risk you're dealing with if something is very risky and the consequences of getting that particular answer wrong or making mistakes there is high then you always have to make sure like you verify it manually let's say this is your master's thesis for your graduate studies and or and you you asked AI to proofread it and they said okay this is done well you got to check it yourself before you submit because it's very risky for you yeah but on the other hand if it is not very risky then probably you're sending a like a long conversation like a long message to your friend and then you want it to fix some typos in there yeah well sure if you make if it makes some mistakes there you can always say hey sorry you know it is an autocorrect or whatever yeah no that makes sense um so i got a i guess a very simple question compare Gemini 3, Chai GPT and Groke which one is the best um so the way i use these models uh i i never say that a single model is the best model for everything so each have their own different strengths the primary reason being that these models are trained on different kinds of data and different ways of training okay and uh pretty much no lab none of this research let's talk about each of them what kind of data or sources they are trained from the majority of the data is going to be publicly available data um like wikipedia a lot of books uh pretty much all the information on the internet like blogs but you said each of them is different though a little bit different so that's the common information which is on the internet publicly available uh and also these models have access to proprietary information which in some cases are not accessible by others uh youtube for example they have access to the whole catalog of videos on um sorry google yeah they have access to youtube right they can use that for their models without any issues but others they have to literally scrape the website and that's not very easy and youtube has terms against you know scraping and all these things so for the others they won't have access to that data and uh this model makers also have uh private deals with uh with many data providers now reddit stackoverflow and news sites like epoch times maybe or like new york times and all these places so they have deals with them to get access to their news and all the new information they publish and so since this they have different deals with different providers the data which the model is finally trained with is kind of varying across the providers yeah like grok is probably to have the full access yeah grok has full access to x or twitter right uh which is why it'll be able to uh in terms of knowing what's really happening in the world right now because pretty much all the news breaks on the internet on x right people so grok would be able to will be having an advantage in that case right uh what about open ai because it doesn't at least to me it doesn't have its own um like a platform or the chat platform so does it have any special prioritary data yeah i think they they have uh some deals with some news providers to get access to their uh news uh basically their data um i don't remember the names i think they were pretty much the very first companies who made such kind of deals exclusive well i don't think it's exclusive but then deals with some news providers i see and where do you see these models going um so there's a trend that bigger models are smarter um so that's something in terms of like you know uh something people have found that there is no limit to that yet uh so that is something they're going to do uh i think gemini is probably the biggest of all the models which are out there and people will try to make even bigger models uh then then the other thing is data so these models need a lot of data and pretty much they've exhausted all the publicly known data so now what kind of proprietary data they can get is is going to be one new thing and people are also exploring uh synthetic data which is again data generated by a model used to train a model so those are some directions in which they're going to use uh to train these models um in terms of how they might evolve uh one direction would be uh now now this since these models are very capable now people don't want to use the models just for a simple task or they wanted to uh be able to do more more complex tasks more independently like like agents like you would have heard the word agents everywhere some called year of agents year of that and whatnot so agent is um like in in one way it is it is an llm or a model like this which has access to uh tools which it can call uh and then probably it'll call a bunch of tools and then loop and then do a repeat that a bunch of times to achieve a particular goal uh what i mean by calling tools is that this model uh can will tell you hey uh the model will do normally something like a thinking process when you do it when i give it a task the model is going to a thinking process and then it'll figure out oh as part of the thinking uh i need to then come probably with a plan and then it needs to exclude the first step of the plan and the first step of the plan is uh to find some information and then it'll realize oh i don't have access to this information um so which is why it'll tell that hey i need this particular information and i can tell it what tools it can use and one tool might be a search tool and i'll tell it how exactly like a detective agent yeah so it's going to uh give uh give me the tell me hey i need you to call this particular tool with this particular uh with giving passing this particular input to that tool and give me the output of that tool and so the person who is calling that llm which is me i'm going to call this tool and then get the information and pass it to the model and then the model will use that information to continue the next step so that is how they use tools in a sense it's like they ask you for like you give a list of tools and the capabilities of that tools and then as part of executing that plan it will figure out what tools i need to use and what data i need to give this input i need to give for these tools to get the desired output and then the person who's calling the tool uh is going to make those call those tools and then get the results and then send back to the model so that truly sounds like a an agent like a maybe not a sec or assistant that's sitting next to you in the office and then you ask it to do something for me and then it can complete the whole task and bring it to you and maybe what i need to do is just yes that is that is the that's the value of many of the ai models especially in terms of uh in coding work at least uh that's what uh that's the big thing right now uh like being able to agentically solve a coding related problems especially these days uh so i think down the road they're going to become be able to uh solve these age problems like in the agentic manner uh in a far better way because you'll notice that as this model more and more capable we wanted to give it more and more complex tasks yeah and every time you delete um some of the fundamental constraints still exist as in as the conversation gets longer and longer the model will kind of get a little dumber and dumber yeah so uh so which means that if you if you give it more and more complex tasks now the problem again exists right uh and also these models nowadays are thinking models which means that the model generates a bunch of text as part of the thinking process and then it executes the actual step or plan or gives the answer so thinking also means that it's generating more text again and every time it answers a new question all the previous answers and the previous thinking is generally there which means that the model context will get used up even more faster so it runs into the same problem where its context is limited and then as the class gets complicated they they need to figure out like what to do with the context and that is for longer and more complex things the models still fail pretty badly in a lot of cases that's something i think the the newer generation of models will improve do you think um are they trying to make it more humanized uh for some specific applications yes uh because people i think they use them for role play and things like that uh like some people want it to act as a therapist or something like that and then some people maybe people even just talk to them like the ai models yeah they want them to they want them as a friend like a digital friend and all that um so i think for for cases like that yes uh but generally uh i think those cases are more i would say niche i would say rather than like the mainstream use cases uh these companies would just focus on the mainstream use cases because that's the one that's going to make them a lot of yeah let me ask you how much in general percentage wise um for an engineer use ai model in terms of coding uh well i think these days a lot of people would you say like 50% of the codings are written by ai uh that depends on uh you know again depends on the organization themselves uh i find that the code generated by these models are pretty reliable these days uh they don't contain a lot of bugs uh and then again as if you're able to use the models in a smarter way you know break down the task and don't give it like too complex tasks and things like that you should be able to give get very good quality code from them um so depending on how skilled the person is in using the models uh they'll be able to get a more a lot of value out of it in terms of how much people use it i don't know i can't tell uh but i'm pretty sure it's uh it's more than ever before yeah i know and it's only getting more yeah it's only yes it's going to be more and more people yeah and we talked a lot about the resources um the ai models used for training purposes etc and even like the by every day people been telling the ais what's going on with their lives probably which they probably shouldn't how do you think our privacy of data use can be protected and how can users stay safe um so you can be pretty sure that anything you put on the internet is going to be used to train the models okay so just don't tell them yeah so if you want to be 100 sure that the your data is not going to be used to train a model and don't put it on the internet that's your safest way uh anything else there's a chance that the models are going to use it uh and there's a lot of proprietary data that this this model makers probably use uh so there's a lot of gray area in there so some proprietary information they the most likely they cannot use for example something you put on your google drive or something like that uh well they just cannot use a data frame models uh but then in some other areas it becomes a gray areas there uh so now let's say you're an author and then your book there's there's some copy of your book on on the internet so will they use it or not well that becomes part of the gray area now yeah so that is one aspect of uh you know data so as long as something about you is on the internet uh you can be sure that it's going to be used yeah in terms of uh this chat models they they they want to use your data uh especially you know the more data they have the better it's you know to you better they're learning and chat gpt or gemini they pretty much have settings in in those chat applications where they say that they can use your data to can we turn it off though and yeah you can generally turn them off uh but then i'm not sure if by default it is on now uh earlier it was on yeah in the very early days of chat gpt i believe then they made the default as off um i heard someone telling that it's on by default now i'm not sure yeah um but yeah you should i think that comes to the ai regulation um what do you think is the most urgent rule for public safety in terms of ai regulation um so like my take is that it's really hard to tell how hard to put in good rules the reason being that the whole ai landscape is moving really fast and changing fast and generally the problems don't happen at uh the first order problems the first order problems are more like consequences of something and the second order problems are the consequences of the consequences can you give us an example uh yeah so uh there's this very famous example called the cobra effect or so basically what happened was um so this was the time when uh when it was when i think the british were ruling over india at that time they found that uh there were a lot of uh cobra snakes in india and they gave people uh an incentive to uh they wanted people to help you know reduce the amount of cobras in there and they said that whenever whoever brings us a cobra head you know we're going to pay you some some amount of money and then they thought this was going to reduce the number of cobras and that was the first order effect they thought that okay we give incentive for people to you know bring the heads of the you know killed cobras and that they're going to kill the cobras and bring just the heads and then it's going to reduce the cobra population then what people figured out was that that oh this is a good way to make money and what they did was they started breeding cobras and then they they started you know killing the cobras and then started giving those heads out and you know making money out of it so that was the second order effect now there's a funny third order effect so once the once the british found that oh this you know people are like scamming us through this process they were like they stopped the whole process and the third order effect of that was that now people had a lot of cobra snakes and they don't know what to do they left them in the wild so now the cobra population increased oh my so that was the third order effect of it that just always shows like the profit is driving everything yeah money arrives a lot of things um so uh i think in the case of ai it's going to be the same kind of things uh like tiktok is a very good example in this case uh so uh like tiktok is very good at uh targeting what exactly your need is and then finding the content relevant to that to that need and getting into rabbit holes and then you're getting addicted on that uh so so what do you think the morality plays a role here uh that again uh so if you if you look what a content you're looking for it gives you more of that content and it gets more and more niche so i don't think it takes morality as a consideration there so if you're looking for bad content it gives you more of that bad content probably uh if you're looking for good content you'll give more of that good content so morality is not a component there now that becomes a problem right yeah and also this uh the uh like the whole app and the way the things the features and the way the app is the trending things and all that in the app they're all very addictive which means that uh like people really none of these kinds of things existed uh even a few years back well when i said a few years back a decade or two decades back so since uh we really don't know what the long-term effects of all these things are since we're experiencing all of these things right now yeah so it'd be very hard to tell and regulate many of these things because you really don't know what the long-term effect is and which is why um so people when they use tiktok they refer to the word called brain rot which basically means that you're kind of getting dumber using tiktok yeah it's like you have a lower attention span you know all these kinds of other uh cognitive problems right but then people only started noticing this what uh two or three years back or something like that i don't think the word brain rot existed two or three years back or maybe even more recent uh and tiktok has been in operation for i think seven or eight years or maybe even longer so it took so many years for people to realize the affect the cognitive effect of this particular thing right and uh llms and ais are are even bigger uh in terms of the impact they have uh so how these things are going to affect people uh in good ways or bad ways it's really hard to tell uh the higher order consequences like the way we talked about the cobra effect right um so that is why i think like rules and regulations won't help or by the time we realize like these these are the problems it might be a little too late in the sense that would be too late for sure yeah so then let's focus focus a little bit more about the current problem which is already happening because how smart the ai tools has been a lot of people are worried about their employment so uh how would you suggest people to stay employed in this ai age and what kind of like skills or mindsets they should be focused on uh there are still a lot of things which ai cannot do that well um so in many cases uh you might have heard that ai is going to take a software engineer's job uh that was something which people have been telling for some time now yeah but you find that that's not really true what it's actually doing is empowering uh software engineers so like you know that the ai model is not going to completely replace a software engineer do all the things a software engineer can do which means that uh like if you're smart enough you should be able to use the ai model to enhance your productivity so in a sense it is another tool you can use to make yourself more productive so do you do you think you can keep your job in five years i think so that's great i i don't think that uh ai is going to replace people in many of the many of the jobs it's going to most likely in many many of the cases it's going to empower people so that think of it as a tool that can help you become better at your job more productive learn better uh do your job better and that's what people have been seeing in cases of uh the software engineering domain and there are a few things which ai still does not have and which is what uh people like you need still need people for many things uh one good case is uh like taste like ai can generate a lot of code ai can do a lot of things but then when it comes to things like having taste of like what is worth building what is worth doing all the nuances yeah those nuances are something people gain through experience and uh it's people have the intuition about these kinds of things and and many people don't have taste right some people do have some people don't and the people who do have they get it through a lot of work like through doing things and and i don't think none of any of thing that thing will go out will go out so which means that uh so in this way it's more like ai is still your tool you can use and you can get some of the things done way better than you were able to do earlier uh and in addition to that there's also the case where uh in terms of uh these tools think of yourself as orchestrating all these tools uh all these agents or these models uh you still are the ones who have to have to make decisions on a lot of things like is it worth building or uh if these are the ways i'm building this what are the trade-offs what are the consequences yeah so uh having good decision making skills is going to be a crucial skill yeah that people need to have and then the better it's going to be uh like people with good taste people with good decision making skills will be able to leverage a lot of value out of it yeah would you say like ask smart questions or the right questions to ai tool not something like should i marry this guy those big decisions or should i invest like a million dollars in some stocks right that's probably not what yeah those kinds of things i don't think ai can do that um so one of those uh i think i can do a lot of things it's just a matter of if you trust them or not uh again it's the risk right like trust by verify uh but then if it is especially risky thing i really would want to verify what it says and then just because ai says oh this guy is great you should marry him you're not going to marry some guy like that right um so uh yeah these kinds of things i would i generally see that uh using it as a tool to to help me do my work better is is going to be the my main use case and i find that these tools will provide a lot of value to a lot of people in that way that's very well said well thank you so much kishore for your time really appreciate you thank you again for hiring me so gem library is amazing it's like a super smart assistant who never sleeps but with all that power is worth slowing down like buckling up in a self-driving car you're still in charge don't treat it like a diary and always double check its answers it's a tool not a crystal ball it can help a lot but it can't make the big decisions for you until then keep your hands on the wheel and your brain switched on see you next time